A Novel Partial Discharge Diagnostics using CAPD with Neural Network to Identify the Vital Defects in XLPE Cable System under DC Stress

Abstract

The application of the XLPE cable under DC stress is now being considered as one of the promising tool for enlarging the limited transmission capacity of the traditional electric AC power cable due to the several technical advantages. Therefore, many research institutes are trying to improve the performance by carrying out experimental investigations related to the reliability of the cable system to be operated under DC stress. One of them is Partial discharge (PD) detection, which is admitted as a reliable tool to indicate the insulation state of the power apparatus under service at the power grid; however, very few works have been reported since last long years. In this work, in order to understand the DC PD produced from the vital defects introducible into the XLPE Cable system, experimental investigation has been carried out under DC applied voltage at the laboratory. For this purpose, three artificial defects are chosen and fabricated to produce DC PDs: Void, Surface and Corona. The presence of void in the cable insulation is imitated by the cavity introduced into the epoxy in such a way to maintain the same electric field distribution as that of produced in real cable. Surface discharge occurring along with dielectric interfaces at the cable joints or terminals is simulated using a solid insulation disk which is fixed between high voltage and ground terminal. Corona discharge, as the result of high divergent electric field spots, is imitated by use of a sharp steel wire intentioanlly located at the sample cable terminals. For the analysis of DC PD patterns related to the defects, our CAPD method, proposed in 2000, has been adopted by considering mainly two parameters: the magnitude of the PD pulses and the time interval between two consecutive pulses. Each defect has been represented in different way by using three patterns respectively. From each pattern, 2-dimensional matrix is obtained followed by one dimensional feautre vector as a traning data and thus each defect has 3 feautre vectors for being appled to calculate the recognition rate by traing through different neural network techiques. They are Multilayer perceptron (MLP), Self-Organizing Feature Map (SOFM) and Recurrent Network (RN) and finally their pattern recognition rate is compared each other in order to find the best techniques among them. Moreover, in this work, in order to improve the recognition rate of the PD pattern obtained from only CAPD data, a consideration has been proposed. In this concept, one feature vector based on the frequency spectrum data of PD signals has been obtained and combined with CAPD data and then similar training process is performed. Finally the recognition rates of PD defects are compared for both the cases. It is observed that CAPD data combined with spectrum data shows much better recognition rate and less mean square error as compared to those of the former case. Also, the MLP technique has shown the best results among the three NN techniques. In future, this idea can be applied to recognize different PD defects with better recognition rates. This can avoid the power apparatus failure and avoid heavy damages of the society.

|AC XLPE전력케이블의 전송용량을 확대하기 위한 방안의 하나로서 DC XLPE 케이블의 적용이 대안으로 고려되어, 많은 해외 연구기관들이 DC 전압 하에서 사용될 케이블의 신뢰성과 관련된 실험적 연구를 다양하게 수행하여 운용의 신뢰도 향상을 위해 노력하고 있다. 일반적으로 폴리머 절연 전력케이블의 절연상태 진단을 위하여 선호되는 방법 중 하나는 부분 방전 (PD) 검출에 의한 결함원 파악이나, DC 전압에서 운용되는 XLPE 케이블에 대한 연구결과는 거의 보고되지 않고 있다. . 본 논문은, DC전압에서 운용되는 XLPE 케이블시스템에서 발생되는 부분방전을 연구하기 위하여 케이블 시스템에 유입되어 치명적인 사고를 유발시키는 3가지 결함을 모의한 인공 결함, 즉 보이드, 표면 및 코로나 결함들을 설계 및 제작 하였다. 케이블 절연체내 보이드는 실제 케이블에서 생성 된 것과 동일한 전계 분포가 유지될 수 있도록 캐비티를 에폭시 사이에 삽입하여 제작하였다. 케이블 조인트 또는 단말 표면에서 발생하는 표면 방전은 고전압과 접지 단자 사이에 고정 된 고분자 디스크를 사용하고, 높은 전계 집중에 의한 코로나 방전은 시료 케이블 단말 중심에 날카로운 강철 와이어를 삽입하여 제작되었다. 인위적 결함에 의한 DC PD 패턴분석을 위하여, 본 연구실에서 2000년에 제안한, CAPD 기법에 근거하여 2개 파라미터, 즉 부분방전 펄스의 크기와 두 개의 연속적인 펄스 사이의 시간 간격, 들이 고려되었다. 실험에 사용된 결함은 3개의 패턴으로 각각 표현되며, 각 패턴은 2차원의 마트릭스로 전환되어 1차원의 특성벡터가 추출된다. 이러한 방법으로 각 결함으로부터 3개의 특성벡터가 얻어지고 학습을 통해 각 결함과 관련된 패턴들의 인식률 산출에 사용된다. 이를 위하여 사용 된 3개의 신경네트워크 기술은 다음과 같고 각각에 의한 인식률들은 상호 비교되어 바람직한 기술을 제안하였다: MLP (Multilayer Perceptron), SOFM (Self-Organizing Feature Map) 및 RN (Recurrent Network). 본 연구는 결함의 인식률을 향상시키기 위해 주파수 스펙트럼 데이터를 CAPD 데이터와 결합시키는 방안이 제안되어 특성벡터를 구하여 상기와 같은 과정을 수행하여 단순히 CAPD 데이터만 사용한 방법과 인식률을 상호 비교하였다. 결과적으로 CAPD와 스펙트럼 데이터를 결합한 방안이 낮은 평균 제곱 오차와 더 나은 인식률의 측면에서 이전 사례와 비교하여 훨씬 더 우수한 성능을 보인 것으로 나타났다. 또한 MLP 기술은 세 가지 기술 중에서 가장 좋은 결과를 나타냈다. 향후, 이 기술은 부분방전 검출과 결함 패턴의 인식률 향상에 적용되어 전력기기의 절연상태 파악 및 원인 분석을 실시간으로 수행할 수 있는 가능성을 제시한 것으로 사료된다.; however, very few works have been reported since last long years. In this work, in order to understand the DC PD produced from the vital defects introducible into the XLPE Cable system, experimental investigation has been carried out under DC applied voltage at the laboratory. For this purpose, three artificial defects are chosen and fabricated to produce DC PDs: Void, Surface and Corona. The presence of void in the cable insulation is imitated by the cavity introduced into the epoxy in such a way to maintain the same electric field distribution as that of produced in real cable. Surface discharge occurring along with dielectric interfaces at the cable joints or terminals is simulated using a solid insulation disk which is fixed between high voltage and ground terminal. Corona discharge, as the result of high divergent electric field spots, is imitated by use of a sharp steel wire intentioanlly located at the sample cable terminals. For the analysis of DC PD patterns related to the defects, our CAPD method, proposed in 2000, has been adopted by considering mainly two parameters: the magnitude of the PD pulses and the time interval between two consecutive pulses. Each defect has been represented in different way by using three patterns respectively. From each pattern, 2-dimensional matrix is obtained followed by one dimensional feautre vector as a traning data and thus each defect has 3 feautre vectors for being appled to calculate the recognition rate by traing through different neural network techiques. They are Multilayer perceptron (MLP), Self-Organizing Feature Map (SOFM) and Recurrent Network (RN) and finally their pattern recognition rate is compared each other in order to find the best techniques among them. Moreover, in this work, in order to improve the recognition rate of the PD pattern obtained from only CAPD data, a consideration has been proposed. In this concept, one feature vector based on the frequency spectrum data of PD signals has been obtained and combined with CAPD data and then similar training process is performed. Finally the recognition rates of PD defects are compared for both the cases. It is observed that CAPD data combined with spectrum data shows much better recognition rate and less mean square error as compared to those of the former case. Also, the MLP technique has shown the best results among the three NN techniques. In future, this idea can be applied to recognize different PD defects with better recognition rates. This can avoid the power apparatus failure and avoid heavy damages of the society.